MMEFU-Net: A Mamba-Guided Multi-Encoder Fusion U-Net for Tumor Segmentation in CT Images
Renzheng Xue, Zifeng Zhang, Yaxin Zhao, Qing Zhang, Minghui Liang · IEEE Access · 2025
Tumor segmentation is crucial for cancer diagnosis, treatment planning, and surgical interventions. However, traditional manual delineation methods are time-consuming, labor-intensive, and prone to subjective variability, highlighting the need for automated approaches. The significant variability in lesion shape and size within medical images further complicates segmentation tasks, necessitating models that can effectively capture both local and global features to achieve accurate results. To address these challenges, we propose MMEFU-Net, a novel deep learning framework designed for efficient and accurate tumor segmentation in medical CT images. The architecture integrates a multi-encoder structure comprising a detail branch, a context branch, and a Mamba-guided branch to extract complementary features, addressing the limitations of single-encoder models. Key innovations include the Mamba-Guided Fusion Attention (MGFA) module for balancing local details and global context, the Pixel Attention Feature Fusion (PAFF) module for enhancing detail branch features with complementary information from the context branch, the Dilated Multi-Scale Fusion (DMSF) module for robust multi-scale feature modeling, and the introduction of the lightweight DySample operator for precise and efficient dynamic upsampling. Comprehensive experiments on five public and private benchmark datasets, encompassing both liver tumors and lung cancers, demonstrate that MMEFU-Net outperforms state-of-the-art U-Net, Transformer-based, and Mamba-based architectures. The model achieves superior Dice Similarity Coefficients (DSC) and Intersection over Union (IoU) scores while significantly reducing computational costs. Notably, MMEFU-Net improves the DSC by 2.16% compared to nnU-Net on the LiTS2017 dataset, with a 35× reduction in parameters and a 25× reduction in computational complexity. These findings suggest that MMEFU-Net has the potential to become a scalable and efficient solution for tumor segmentation in clinical practice.